Fix chrF fallback so English wiki garbage fails translation_success

Sacrebleu errors used to fall back to set-overlap unigrams, so any
English hyp scored ~70–80 against English refs and SFT reported
success 1.0. Use count-based char n-grams, keep BLEU failures from
clobbering chrF, and print a few hyps during eval.

tests/integration/test_eval_mt.py

@@ -28,6 +28,16 @@ def test_good_zh2en_passes():
     assert translation_success(src, hyp, ref, target_lang="en") is True
 
 
+def test_english_wiki_garbage_does_not_pass_zh2en():
+    from kda.training.success import _chrf
+
+    src = "今天天气很好。"
+    hyp = "The first one's the time."
+    ref = "The weather is very nice today."
+    assert _chrf(hyp, ref) < 40.0
+    assert translation_success(src, hyp, ref, target_lang="en") is False
+
+
 def test_container_help_exits_2():
     import importlib.util
     from pathlib import Path